Frequent constraint violations on the data stored in a database may suggest that the semantics of the represented reality is changing. In this work we propose a methodology and a tool, based on data mining, to maintain the integrity constraints specified at design time, in order to adjust them to the evolutions of the modeled reality that may occur during the database life. The approach we propose allows to isolate frequent and meaningful constraint violations and, consequently, to extract novel rules that can be used to update/relax the no longer up-to-date integrity constraints.